Metastatic tumors, which are cancerous cells known for breaking away from their tissue of origin, move through the body via the lymph or circulatory systems, and eventually create new tumors in other body parts, are highly difficult to detect.
In fact, a 2009 study involving 102 breast cancer patients, which was carried out at two Boston health centers, revealed that one out of four patients were affected by the failures of the “process of care” like incomplete diagnostic tests and insufficient physical examinations.
This forms one of the main reasons that of the 500,000 deaths globally caused by breast cancer, nearly 90% were the outcomes of metastasis. However, researchers at Google AI and the Naval Medical Center San Diego have created a promising solution that applies cancer-detecting algorithms that independently go through the biopsies of lymph nodes.
Their AI system, Lymph Node Assistant, is described in a paper referred to as “Artificial Intelligence-Based Breast Cancer Nodal Metastasis Detection,” which featured in The American Journal of Surgical Pathology. When tests were conducted, it attained an AUC or area under the receiver operating characteristic of 99%.
“Artificial intelligence algorithms can exhaustively evaluate every tissue patch on a slide. We provide a framework to aid practicing pathologists in assessing such algorithms for adoption into their workflow (akin to how a pathologist assesses immunohistochemistry results),” wrote the authors of the paper.
LYNA, based on Inception-v3, is an open source image recognition deep learning model that has been said to attain more than 78.1% accuracy, particularly on Stanford’s ImageNet dataset. According to the researchers, it takes a 299-pixel image as input, identifies tumors at the pixel level as well as during training.
The team made an improvement on algorithms that had been published in the past by exposing the LYNA to a 4:1 ratio, primarily that of normal to tumor patches as well as boosting the training process’ “computational efficiency”.
The researchers employed LYNA to the 2016 Cancer Metastases in Lymph Nodes challenge dataset, which is a collection involving 399 whole-slide lymph node image sections drawn from the University Medical Center Utrecht and Radboud University Medical Center.
In tests, LYNA attained a slide-level accuracy of 99.3%. After the sensitivity threshold sensitivity was changed to detect all the tumors on each slide, it showed 69% sensitivity, by accurately spotting all the 40 metastases present in the evaluation dataset devoid any false positives. What’s more, it was unaffected by the slides’ artifacts including overstaining, hemorrhage, poor processing, and air bubbles.
LYNA was not perfect. In fact, it often misrecognized histiocytes (bone marrow-derived white blood cells), germinal cancers and giant cells. However, it was able to perform better compared to a practicing pathologist assigned with the role of evaluating similar slides.
“[Lyna] achieves higher tumor-level sensitivity than, and comparable slide- level performance to, pathologists. These techniques may improve the pathologist’s productivity and reduce the number of false negatives associated with morphologic detection of tumor cells,” wrote the researchers.
DeepMind, which is Google’s AI research division in London, takes part in several health-based AI projects. Previously, it collaborated with the UK-based National Health Service to build an algorithm that could find the early blindness signs.